As the population grows and arable land decreases,crop yields become more and more important.Applying deep learning to the classification of crop diseases can timely and accurately detect diseases,which is of great significance for improving crop yields.There are two datasets in this paper.One is a rice dataset,with about 500 images,10 classifications(10 common rice diseases,this paper investigated the symptoms and treatment measures),all of which were taken in real farmland environments.The other is the PlantVillage public dataset,with a total of 54306 leaf images,of which there are 14 plants and 38 classifications(including healthy samples).Aiming at the rice disease image data set with complex background conditions,this paper proposes a new image segmentation process based on Laplace edge detection operator,morphological method and OTSU.The steps are as follows:1.Convert the original image into a grayscale image;2.Use the edge detection operator to process the gray image of the original image;3.Use morphological method to process;4.Use OTSU method for front and back scene segmentation;5.Draw the contours;6 Screen out the contours with the largest circumscribed rectangle;7.Mask the original color image to get the final result.Experiments show that this algorithm can effectively extract the lesions of rice from complex backgrounds,and finally improve the classification accuracy of models.This paper uses a two-stage training method to perform transfer learning on 5 improved models on 2 data sets.By comparing the results,we have got the following conclusions:1.The two-stage training method has advantages in improving and stabilizing accuracy;2.InceptionResNet-v2 achieved the highest test set accuracy of 94.93%on the rice dataset.Performing transfer learning on InceptionResNet-v2 is most likely to succeed.Although MobileNet-v2 is not stable enough,the model has the smallest volume,the shortest training time,and high accuracy.MobileNet-v2 achieved the highest test set accuracy of 99.74%on the PlantVillage dataset.MobileNet-v2 is the model with the best applicability;3.NASNetMobile is most sensitive to the size of the data set.NASNetMobile requires large data sets and longer training time to get good results.This paper also compares the results of other papers using the PlantVillage public dataset.Compared with most of the results,this paper still obtains a higher accuracy with a larger number of classification categories.DenseNet’s accuracy is only 0.01%higher than MobileNet-v2,but MobileNet-v2’s volume and model depth are smaller than DenseNet’s,and its applicability is stronger.This paper also made a comparison experiment of transfer learning,which proved that the use of pre-trained models and model improvements can accelerate model convergence under the same training algebra.Using pre-trained models with model improvements can result in higher accuracy.Finally,this paper designs and implements a pest identification system by combining the improved model. |